www.ICA2014.

org

A Primer in Multilevel Modeling

for Actuarial Applications

Mona Hammad, ASA PhD

Assistant Professor, Math & Insurance dept.,
Faculty of Commerce, Cairo University, Egypt.
mona_hammad@hotmail.com
 Agenda
• Overall aim of the paper

• Motivation

• Overview of key points

• Conclusion

• Next steps

2
 Overall Aim of the Paper
• a simple introduction

• suitable for both academics and

practitioners

• in order to encourage more to use this

important technique in their future work.

3
 Motivation
• Hammad, Mona S. A., & Harby, Galal A. H.
(2013). Using Multilevel Modeling for
Group Health Insurance Ratemaking: A
Case Study from the Egyptian Market.
Paper presented at the Perspectives on
Actuarial Risks in Talks of Young
Researchers (PARTY2013), Ascona,
Switzerland.
www3.unil.ch/wpmu/party2013/
4
 Overview of key points

1. What is multilevel modeling?

2. Why to use multilevel modeling?

3. When to use multilevel modeling?

4. How to use multilevel modeling?

5
 1- What is multilevel modeling?
• multilevel model can be defined as:
– a model “specified in stages, with each stage
building upon another” (Searle, Casella, &
McCulloch, 2006, p. 315).

– “a regression (a linear or generalized linear

model) in which the parameters—the
regression coefficients—are given a
probability model.” (Gelman & Hill, 2007, p. 1)

6
Simple Illustrative Example

• dataset of observations for units i nested

groups j. (examples?!)
• Hierarchical/
clustered/ Nested
data structures.
Group j
•  Reason for
alternative name
HLM / limitation

• More layers
Unit i
• Panel/longitudinal
data 7
Simple Illustrative Example (Cont.)

No Pooling
Complete Pooling
(separate models)
(one single model)
Not practical
(-) ignore &
correlation Partial Pooling
(-) not suitable for (Multilevel Model)
all? a compromise

single model but allows for heterogeneity

between subjects and groups in the
dataset. (Gelman & Hill, 2007)

8
A] Varying/Random B] Varying/Random C] Varying/Random
intercept model slope model intercept and slope
model

Note: each of these graphs assumes a hypothetical dataset with 6 groups (i.e. six level 2 units), with each
fitted line representing a different group, in other words. Source: (Gelman & Hill, 2007, p. 238)

Possibility of different explanatory

variables at different levels
9
 2- Why to use multilevel
modeling?
• allowance for:
– heterogeneity between different subjects in
the dataset without the need to fit separate
models.
– possible correlations over time between
observations related to the same subject
and/or group.
It can handle
complex data
structures
10
 3- When to use multilevel
modeling?
• Complex data structures such as:
– Hierarchical data

– Panel data

– Cross classified data

11
 4- How to use multilevel
modeling?
• Specialized software vs. general purpose
statistical packages (with mixed model
procedure).
Hierarchical Linear Model (HLM) Linear Mixed Model notation
Level 1: yij   00  r0 j  eij
 
Fixed Random
Level 2:

12
 Conclusion
• Powerful tool

• Open a wide spectrum of potential model

designs
– need for careful planning (i.e. a clear
modeling strategy).

13
 Next Steps
• Read the paper?

• Send feedback?

g{tÇ~ lÉâ
14
 References
• Gelman, A., & Hill, J. (2007). Data Analysis
Using Regression and
Multilevel/Hierarchical Models. New York:
Cambridge University Press.
• Searle, S. R., Casella, G., & McCulloch, C.
E. (2006). Variance Components. New
Jersey: John Wiley & Sons, Inc.

15